Referring to FIG. 1, a conventional system 10 for implementing multiqueue FIFOs is shown. The system 10 generally comprises a selector section 12, a selector section 14 and a number of memory sections 16a-16n. The memory sections 16a-16n are implemented as FIFOS. The conventional system 10 implements each of the FIFOs 16a-16n as an independent physical memory.
The selector section 12 receives data from a write interface and presents the data to one of the memory sections 16a-16n in response to a write select signal WR_SEL. The selector section 12 selects one of the FIFOs 16a-16n based on the signal WR_SEL. The incoming data is then stored into the appropriate FIFO 16a-16n. Similarly, the selector section 14 presents data to a read interface from one of the memory sections 16a-16n in response to a read select signal RD_SEL. The selector section 14 selects one of the FIFOs 16a-16n based on the signal RD_SEL and reads the data from the appropriate FIFO 16a-16n.
Independently implementing each FIFO 16a-16n as a separate memory element is inefficient. Additionally, the conventional system 10 does not provide flexibility for implementing variable sized FIFOS, since an increase of each FIFO in depth can only be achieved by cascading two or more of the FIFOs 16a-16n. Such conventional depth expansion of the FIFOs 16a-16n occurs in large quantities and can be inefficient (i.e., depth is only increased by implementing an additional FIFO). Additionally, increasing the size of the FIFOs 16a-16n reduces the number of FIFOs that can be implemented. Moreover, the conventional system 10 is not scalable for a large number of FIFOs.
One implementation of the conventional system 10 stores multicast port information into one of the FIFOs 16a-16n in response to the incoming data. However, with such an approach, one particular FIFO 16a-16n of the system 10 can become significantly large, based on memory density and packet size of incoming data. For example, to implement twice a minimum size packet, half of a port information memory will be wasted. For even larger size packets, port information memory utilization is significantly reduced.
If the memory size is 64K times the data width and the minimum packet size is 8 times the data width, the number of minimum size packets is 8K (memory size/minimum packet size). For each multicast packet, port information is stored. Assuming all of the 8K packets are multicast, a 8K×32 memory is required to store all the port information for the 8K multicast packets (i.e., ×32 for a 32 port configuration or ×16 for a 16 port configuration). Maximum memory utilization occurs when all 8K are multicast packets and each is of minimum packet size. If packet size is increased, assuming all are multicast packets, the memory utilization will vary as follows:
Packet size (words)% mem utilization8100%165032256412.51286.25
If packet size is minimum, and multicast packets are 10% of the total number of packets, then memory utilization will be 10%. Similarly, for a 64 word packet size, if 10% of the total number of packets are multicast, then the memory utilization is 10% of 12.5=1.25%.
Another implementation of the conventional system 10 implements separate logic structures to utilize unicast and multicast address generator logic. However, such an approach requires significant additional logic and has a limited amount of logic reuse. Furthermore, such an implementation can have high power consumption.
To change the depth of the FIFOs 16a-16n of the conventional system 10, a queue pointer memory data can be stored in a storage element (not shown). However, a default value for each depth field of the FIFOs 16a-16n needs to be stored. During power up of the conventional system 10, the default value is stored in the depth field for each of the FIFOs 16a-16n.
With such an approach, the dual port memory is configured to store the queue pointer memory value and is written on power up even when the default values are to be used. Such writing during power up requires extra cycles during initialization. Additionally, for a large number of FIFOs 16a-16n, a large number of initialization cycles is required.
Conventional multiqueue memories use a forward pointer memory to store a link-list for creating virtual queues in the main memory. The forward pointer memory implements pointers to the queues in the main memory. During power up the forward pointer memory is required to be initialized in order to ensure that the entire memory is linked and is available for use. In such a conventional method, during configuration, all the locations of the forward pointer memory are initialized.
Furthermore, the forward pointer memory has to be written to on power up to initialize the link-list, which requires a significant configuration time. Additionally, the configuration time of the forward pointer memory is dependent on the size of the forward pointer memory and increases linearly as the size of the forward pointer memory increases.